Claude API Cost in TWD 2026: Opus 5, Sonnet 5, Haiku 4.5

Anthropic prices in USD, but Taiwanese companies often plan budgets in TWD. This table converts the USD unit prices of Claude's main models (Opus 5, Sonnet 5, Haiku 4.5) into NT$, and adds monthly cost estimates for three real scenarios so you can see actual spending.

Bottom line up front: Sonnet 5 is the default choice for most production tasks ($2 / $10). Haiku 4.5 suits high-frequency simple tasks (its input price is only half of Sonnet 5's). Save Opus 5 for the most complex agent workflows (2.5 times Sonnet 5's price).

⚠️ If you are still using Sonnet 4.6, switch directly to Sonnet 5. The two serve the same tier, but Sonnet 5 is $2 / $10 and 4.6 is $3 / $15. It is newer and cheaper, so there is no reason to keep the older version. Similarly, Opus 4.5 through Opus 5 are all priced at $5 / $25, so choose the latest one. Opus 4 and Opus 4.1 are $15 / $75, three times the price of the newer models; they are only worth keeping if they are hard-coded in existing code and you do not want to change them for now.

Update (2026-10-08): Claude Haiku 5.5 is now live at $0.10 input / $0.50 output (per million tokens), with a 1,000,000-token context window. For a single request with more than 100,000 input tokens, the entire request is charged $0.50 / $2.50. This is lower than Haiku 4.5 below ($1 / $5), so the cheapest current Claude is Haiku 5.5. This article was written before Haiku 5.5 was available; see Should you switch from Claude Haiku 4.5 to 5.5?. A Haiku 5.5 row has been added to the price table below.

Update (2026-09-29): Claude Sonnet 5.5 is now live, priced the same as Sonnet 5 ($2 / $10, 1,000,000-token context window). Read the "Sonnet 5 is the latest" statement above as "Sonnet 5.5 is priced the same as Sonnet 5, so new projects should evaluate Sonnet 5.5 directly." The Sonnet 5 cost estimates in this article also apply to Sonnet 5.5. For details, see Should you switch from Claude Sonnet 5 to 5.5?.

Update (2026-09-23): Claude Opus 5.5 is now live, priced lower than Opus 5 ($4 / $20 vs $5 / $25). This article was written before it was available. New projects should evaluate Opus 5.5 directly. The Opus 5 estimates in this article remain useful for existing projects already running Opus 5. For details, see Should you switch from Claude Opus 5 to Opus 5.5?.

Want to skip buying directly and pay in TWD with a unified invoice for your company? BazaarLink provides an OpenAI-compatible route to Claude at the same model prices, plus a 10% top-up service fee, and it solves the expense reporting pain points for Taiwanese companies. For details, see Complete guide to expense reporting for AI APIs in Taiwan.

Model cost comparison (USD ↔ NT$)

The exchange rate is 1 USD = NT$ 31.5 (2026 Q2 average). MTok means million tokens. The "multiple of Sonnet 5" column uses Sonnet 5 as the baseline for cross-model comparison.

These unit prices match the USD prices shown on the BazaarLink models page, with no model markup (priceMultiplier = 1.0). For the latest prices, see BazaarLink Models.

If you need to purchase under a company name and require a tax ID and unified invoice, first read How to fill in a company tax ID for Claude? Can you get a unified invoice in Taiwan?.

Three real-world scenarios (NT$/month)

Scenario 1: Customer service bot (Sonnet 5, 500 conversations/day)

Assumptions: 500 conversations per day, an average of 3 turns per conversation, and each turn has 300 input tokens / 150 output tokens (excluding repeated system prompt costs; if the system prompt exceeds 1k tokens, enabling prompt caching is recommended).

• Monthly conversations: 500 × 30 = 15,000 conversations
• Monthly turns: 15,000 × 3 = 45,000 turns
• Monthly input tokens: 45,000 × 300 = 13,500,000 (13.5 MTok)
• Monthly output tokens: 45,000 × 150 = 6,750,000 (6.75 MTok)

Cost:

• Input: 13.5 × NT$ 63 = NT$ 851
• Output: 6.75 × NT$ 315 = NT$ 2,126
• Total: NT$ 2,977/month (≈ USD 95)

(With the same assumptions, the older Sonnet 4.6 costs NT$ 4,476/month. Switching to Sonnet 5 saves NT$ 1,499 per month, a 33% reduction.)

Scenario 2: Document processing (Haiku 4.5, 100 documents of 5,000 characters/day)

Assumptions: 100 documents per day, each averaging 5,000 characters ≈ 6,500 tokens, with Haiku producing an 800-token summary per document.

• Monthly documents: 100 × 30 = 3,000 documents
• Monthly input tokens: 3,000 × 6,500 = 19,500,000 (19.5 MTok)
• Monthly output tokens: 3,000 × 800 = 2,400,000 (2.4 MTok)

Cost:

• Input: 19.5 × NT$ 32 = NT$ 624
• Output: 2.4 × NT$ 158 = NT$ 379
• Total: NT$ 1,003/month (≈ USD 32)

→ At the same volume, Sonnet 5 costs about 2× as much as Haiku: NT$ 1,985/month; Opus 5 costs about 5× as much as Haiku: NT$ 4,972/month. For high-frequency document processing, Haiku 4.5 is still the obvious choice.

Scenario 3: Agent workflow (mixed Opus 5 + Sonnet 5, 50 runs/day)

Assumptions: 50 complete agent tasks per day. Each task includes 1 Opus 5 planning call (2,000 input + 800 output tokens) and 4 Sonnet 5 execution calls (1,500 input + 600 output tokens per call).

• Monthly tasks: 50 × 30 = 1,500 tasks
• Opus: input 1,500 × 2,000 = 3 MTok, output 1,500 × 800 = 1.2 MTok
• Sonnet: input 1,500 × 4 × 1,500 = 9 MTok, output 1,500 × 4 × 600 = 3.6 MTok

Cost:

• Opus input: 3 × NT$ 158 = NT$ 473
• Opus output: 1.2 × NT$ 788 = NT$ 945
• Sonnet input: 9 × NT$ 63 = NT$ 567
• Sonnet output: 3.6 × NT$ 315 = NT$ 1,134
• Total: NT$ 3,119/month (≈ USD 99)

→ Running the same tasks entirely on Opus 5 costs NT$ 5,678/month (1.8 times as expensive). The mixed strategy saves NT$ 2,559 per month. For cost-sensitive tasks, the classic approach is to use Opus for planning and Sonnet for execution.

BazaarLink vs buying directly from Anthropic

Claude model unit prices are the same on both sides (no model markup, priceMultiplier = 1.0). The differences lie in the billing currency and the invoicing process:

For Taiwanese businesses, the 10% top-up fee saved by buying directly from Anthropic often has to be made up several times over in accounting, exchange losses, and expense reporting processes.

→ For the full comparison, see How to handle AI API expenses for Taiwanese companies? 2026 complete guide.

Conclusion

Need to use the OpenAI-compatible SDK to connect to the Claude API, with TWD visibility and unified invoices? The BazaarLink free tier (free models at 50 requests per day and 10 requests per minute, no credit card required) lets you test latency and response quality before deciding. For the latest prices, see BazaarLink Models.

TWD billing · Taiwan invoices · leading AI models · OpenAI-compatible API